An Apache-2.0 UK AISI Inspect AI benchmark (N=800) testing silent premise contamination, SI unit cascades, and over-retraction in AI agents.
## Project summary
Long-running AI agents keep persistent working memory across dozens of turns while receiving mid-task corrections from tools, webhooks, or users. When an earlier number or unit is retracted 500 mg corrected to 2 g), frontier models often acknowledge the update in prose while quietly keeping stale downstream totals, dropping unit conversions 500 + 2 = 502 instead of 2.5 g), or overturning unrelated prior conclusions.
**Gyeol-Bench** (Apache-2.0) is an open, standalone benchmark built for **UK AISI Inspect AI** to measure this exact failure mode—**Silent Quantitative Premise Contamination**—across 800 procedurally generated multi-turn tasks. It is strictly a focused public measurement tool and contains zero code from my separate, on-hold private language project Gyeol).
The core benchmark generation and baseline evaluation are already self-funded using my own compute credits. Manifund funding goes strictly toward independent human auditing, public baseline challenge bounties, and cross-lab frontier API runs across Anthropic, OpenAI, DeepSeek, and 70B+ open-weight models.
## What are this project's goals? How will you achieve them?
### Why existing benchmarks miss this gap
Closest belief-revision and multi-turn benchmarks—including **AGM-BENCH**, **ReviseQA**, **BayesBench**, **Belief-R1**, **TRACK (EACL 2026)**, and **FPCO-Dialog (EMNLP 2026)**—test qualitative belief updates, subjective probabilities, or visual false premises. Stateless graders math-verify, Prime verifiers) check single-turn answers. None simultaneously test:
1. Explicit mid-trajectory premise retractions,
2. Multi-hop numerical dependency recalculation,
3. Coupled SI and financial unit/scale changes mg/g/kg, K/M/B, %, bps), and
4. Preserving unrelated earlier conclusions alongside active retractions.
### What I am building
I am shipping a lightweight, clean-room Python package jaerieu/gyeol-bench, Apache-2.0) with four clear parts:
- **Five-condition task generator N = 800, runs locally at zero compute cost):**
- Condition A (400 unperturbed controls): baseline multi-step calculations.
- Condition B: single premise retraction within fixed units.
- Condition C: 3+-hop downstream numerical recalculation after retraction.
- Condition D: simultaneous retraction plus SI or financial scale conversion.
- Condition E: unrelated independent conclusions alongside a retraction, testing whether the agent over-retracts valid results.
- **Independent exact ground truth:** All 800 answers are computed with Python's exact rational Fraction arithmetic (no floating-point drift) and cross-verified by an independent SymPy + Pint script.
- **Four-way error diagnosis:** Every model failure is separated into four distinct buckets—JSON format rejection PARSE_REJECT), arithmetic error given valid format, stale-state retention (keeping the invalidated value), and over-retraction (corrupting unrelated conclusions).
- **One-command Inspect AI plugin & public audit trail:** Runs offline via uv run inspect eval gyeol_bench/retraction_dag under the May 2026 inspect_evals standard. Baseline code and seeds are hash-frozen on Day 46 for a 14-day public PR challenge, daily digests are anchored to **Sigstore Rekor** and **FreeTSA**, and results are hosted on a read-only zero-telemetry portal gyeol.link).
## How will you know it worked?
The project is governed by pre-registered pass/fail gates so work stops cleanly if a core premise does not hold:
- **Day 21 — Technical integrity:** 100% agreement between the Fraction oracle and the independent SymPy + Pint cross-oracle across all 800 items, plus passing unit and zero-logging portal tests.
- **Day 60 — Real difficulty (non-saturation):** Across 3+ held-out open-weight model families and frontier APIs, the strongest model must still fail at least 10% of perturbed tasks. If models already clear >90% even after one harder redraw, the project halts as saturated.
- **Day 60 — Low false alarms:** On the 400 unperturbed control tasks, false alarms must stay at or below 5 / 400 <= 1.25%, 95% upper bound < 5%), with every flagged item manually inspected.
- **Day 60 — Verifier value & human audit:** Any verifier-assisted setup must beat the unassisted baseline (same model, same JSON schema) by at least +5%p with two-sided exact McNemar p < 0.05, backed by 100% manual review of discordant cases and a blinded N >= 150 audit by 2 to 3 external AI safety researchers recruited before Day 46.
- **Day 75 — Open release:** Full Apache-2.0 release of datasets, Rekor proofs, and the inspect_evals validity report.
- **Day 180 — Independent uptake:** At least 3 independent third-party citations/adoptions and at least 1 independent external run; otherwise the repository is cleanly archived with no further funding requested.
## How will this funding be used?
Baseline N = 800 evaluation is already covered by my own compute credits Google Colab 3,200+ CU and Google AI Studio $100 Paid Tier credit). Every grant dollar funds external review, multi-lab API coverage, and reproducibility—with *$0 personal stipend**.
**Budget breakdown (Minimum $6,000 / Max goal $15,000):**
- **External human oracle audit & PR challenge bounties $2,000 min / $4,000 max, Days 15–45):** Compensating independent safety reviewers for the blinded N >= 150 ground-truth audit and offering bounties during the 14-day public baseline challenge.
- **Cross-lab frontier model evaluation $3,200 min / $7,500 max, Days 35–60):** Batch API evaluation across Anthropic Claude, OpenAI o-series/GPT, DeepSeek, and 70B+ open-weight endpoints across v0.1 and a fresh v0.2 seed refresh.
- **Inspect AI v0.2 tool-use suite & reproducibility ops $800 min / $3,500 max, Days 60–120):** Clean-room container verification runs, automated Sigstore Rekor / FreeTSA timestamping, and extending tasks to multi-turn tool-output retractions.
**Overlapping applications:** I am also applying to EA Funds (Transformative AI Fund). Any overlapping compute or audit costs funded elsewhere will be disclosed publicly on this page and offset without duplicate billing.
**Visa compliance & fail-closed payout safeguard:** I live in Brisbane, Australia on an employer-sponsored Subclass 482 visa (working full-time 45 hr/wk in an unrelated hospitality role), governed by Condition 8607 and Migration Regulations 1994 reg 1.03. Because Manifund disburses funds directly to grantees Model C), I have prepared a formal legal inquiry for an Australian registered migration agent (RMA) or immigration lawyer covering grant receipt, third-party Model A 501(c)(3) vendor-direct sponsorship, and unpaid open-source work. *$0 will be withdrawn to any personal account unless written legal advice explicitly clears it.** If personal receipt is restricted, pledged funds will move only if both legal clearance and an approved Model A 501(c)(3) fiscal sponsor paying vendors directly are in place; otherwise 100% of pledged funds stay unwithdrawn or are returned to donors at zero loss.
## Who is on your team? What's your track record on similar projects?
I am **Jae Rieu**, an independent AI safety and epistemic verification researcher based in Brisbane, Australia. Outside my full-time hospitality job, I maintain a strict 10-hour-per-week personal review discipline supported by automated test harnesses and AI coding tools.
Prior completed work (all built at $0 external grant cost):
- **Dual-runtime verification engines:** Built 42,982 lines of zero-dependency TypeScript and Python verification code RFC 8785 canonicalization, arithmetic/unit checks) backed by 978 unit tests and 500/500 cross-runtime parity checks.
- **Apart AI x Epistemics 2026 empirical study:** Completed 240 multi-turn sessions (7,176 API calls) with 18 SHA-256 frozen public reproducibility artifacts, validating high-precision automated audits and isolating small-model syntax failures 11.01%).
- **Public work:** [github.com/jaerieu/Gyeol_Public](https://github.com/jaerieu/Gyeol_Public) and [gyeol.link](https://gyeol.link). Before Day 46, I am recruiting 2 to 3 external AI safety researchers / Inspect AI contributors to run the blinded audit.
## What are the most likely causes and outcomes if this project fails?
- **Visa legal restriction:** If Australian immigration counsel advises that neither direct grant receipt nor Model A vendor-direct payment is permitted under my Subclass 482 visa, $0 is withdrawn and 100% of pledged funds return to regrantors.
- **Benchmark saturation:** Frontier models may already solve >90% of retraction and unit-cascade tasks. If so, I will stop after one harder redraw and publish the negative finding so others do not repeat the effort.
- **Null verifier gain or high false alarms:** If verifier-assisted runs fail to clear the +5%p p < 0.05) bar or controls exceed 5/400 false alarms, I will drop any verifier-superiority claim and publish the raw error breakdown openly.
- **Low community adoption or maintainer bottleneck:** Even a sound benchmark can go unused. I mitigate this via one-command Inspect AI packaging, a tiny clean-room codebase, and early external reviewer recruitment. If fewer than 3 independent teams adopt or cite it by Day 180, I will freeze it cleanly as a static public archive.
## How much money have you raised in the last 12 months, and from where?
*$0 USD** in external grants, donations, prizes, or investment over the last 12 months.
Everything built so far has been 100% self-funded out of pocket (roughly $800 USD in API and hosting expenses, plus pre-owned personal credits: Google Colab 3,200+ CU and Google AI Studio $100 Paid Tier credit). A parallel Stage 1 application is being submitted to **EA Funds (Transformative AI Fund)** with no funds awarded to date.
Comments
No comments yet. Sign in to create one!